CIFAR100-Image-Classification

A collection of modified CNN models implemented in PyTorch for the CIFAR-100 classification task. Models are adapted for $32 \times 32$ inputs while preserving their core architectural logic.
I'm using PyTorch 2.10.0+cu128 in Python 3.12.0.

Structure

├── data/
|   ├── cifar100/
|   ├── processed/
├── models/
|   ├── vgg16.py
|   ├── googlenet.py
|   ├── resnet18.py
|   └── ...
├── config.py
├── utils.py
├── prepare_data.py
├── dataset.py
├── train.py
├── predict.py
└── eval.py

Read files in order:

config.py → utils.py → prepare_data.py → dataset.py → model → train.py → eval.py → predict.py

Requirements

matplotlib==3.10.8
numpy==2.4.4
Pillow==12.2.0
torch==2.10.0+cu128
torchvision==0.25.0+cu128
tqdm==4.67.3

Dataset

The dataset comes from Kaggle website: CIFAR-100. It has a total of 100 categories.
The raw training set has 50,000 images (each category 500 images), the raw test set has 10,000 images (each category 100 images), each image is a $32 \times 32$ RGB image.

CIFAR-100

Data Preparation & Augmentation

Splitting:

To split the data, run the command -

python prepare_data.py

This will split the raw training data into 90% Training (45,000) and 10% Validation (5,000).

Augmentation:

I placed data augmentation in dataset.py, including random cropping, random horizontal flipping and AutoAugment (a reinforcement learning-based strategy that automatically searches for and applies the optimal combination of data augmentation policies). These augmentations can enrich the diversity of the training data, and improve model's robustness and generalization capability. Furthermore, data normalization is applied to stabilize the training process.

Train

To start training, run the command -

python train.py

I used a Cosine Annealing Schedule to adjust the learning rate during training -

scheduler = CosineAnnealingLR(optimizer, T_max=NUM_EPOCHS, eta_min=1e-6)

• T_max: Number of iterations for the learning rate to decrease to eta_min.
• eta_min: The minimum target learning rate.

The learning rate $\eta_t$ is adjusted according to the following formula:

$$ \eta_t = \eta_{min} + \frac{1}{2}(\eta_{max} - \eta_{min})\left(1 + \cos\left(\frac{T_{cur}}{T_{max}}\pi\right)\right) $$

This strategy ensures a smooth transition from the initial learning rate to the minimum.

You can adjust hyperparameters in config.py according to your own hardware (It is recommended to train on a GPU). I used an NVIDIA GeForce RTX 2080 Ti GPU (11GB VRAM).

Prediction

To test your trained model, run the command -

python predict.py

It randomly selects an image from the test set, and displays the image with its label and the model's predicition results (green for correct, red for wrong).

Evaluation

To evaluate your trained model on the test set, run the command -

python eval.py

It will show the model's prediction accuracy on the test set.

Model	Params	FLOPs	Epochs	Top1 Accuracy	Top5 Accuracy
SqueezeNet	776.74K	25.08M	200	68.03%	89.94%
VGG-13	18.95M	295.47M	200	74.55%	92.77%
VGG-16	24.26M	408.85M	200	74.73%	92.16%
VGG-19	29.58M	522.23M	200	74.49%	92.41%
GoogLeNet	6.08M	483.24M	200	78.60%	95.04%
MobileNetV2	2.35M	94.67M	200	71.41%	93.18%
MobileNetV3	4.33M	71.16M	200	73.51%	93.59%
ShuffleNetV1	999.87K	45.36M	200	73.23%	93.27%
ShuffleNetV2	1.36M	47.36M	200	73.54%	93.42%
ResNet-18	11.22M	557.94M	200	78.01%	94.27%
ResNet-34	21.33M	1.16G	200	78.04%	94.91%
ResNet-50	23.71M	1.31G	200	79.35%	95.44%
ResNet-101	42.70M	2.53G	200	79.18%	94.87%
ResNeXt-50	23.18M	1.36G	200	80.36%	95.56%
ResNeXt-101	42.33M	2.59G	200	81.91%	96.18%
DenseNet-121	7.05M	908.19M	200	80.99%	96.01%
DenseNet-169	12.64M	1.08G	200	80.87%	95.72%
DenseNet-201	18.28M	1.40G	200	80.21%	95.26%
WideResNet-40-4	8.97M	1.30G	200	80.05%	95.59%
WideResNet-28-10	36.54M	5.25G	200	81.12%	96.05%
VisionTransformer	4.78M	308.39M	200	61.92%	86.87%